Red-team safety evaluation conducted by Perplexity AI found that 8 out of 10 third-party AI agent sandbox platforms exhibited network-policy bypasses